What OpenAI is building for a post-prompt future

AI is transitioning from just answering questions to doing valuable work. The next challenge is making agents more accessible and simple enough that the technical details fade into the background.

In this episode of The Deep View Conversations, we sit down with two members of OpenAI's ChatGPT Work team, Tara Seshan and Ty Geri, to dig into ChatGPT Work and what OpenAI is doing to make advanced agent capabilities useful to a lot more people. We also dig into some of the current challenges and how the team is approaching them. 

Seshan and Geri explain how scheduled tasks and proactive assistance are changing the way people start their workdays, why AI lets teams move from debating ideas to testing prototypes, and how personalized software can turn one-off needs into purpose-built tools. They also discuss the challenge of token costs and model selection, why "super app" isn't the most useful framing for ChatGPT and Codex, and what it will take for agents to become more persistent, proactive, and connected.

The conversation also covers:
• How OpenAI is trying to bridge local and cloud workflows
• Why Tara and Ty start their days with agents instead of Slack
• Building personal apps and tools without traditional software overhead
• The tradeoff between model capability, cost, and user control
• More persistent agents and proactive personal assistance
• Connecting agents to email, calendars, enterprise systems and third-party tools
• Privacy, security and administrative controls for agentic work

If you’re figuring out where agents fit into your work or what has to improve before you trust them with more of it, then this conversation offers a practical look at how OpenAI is preparing for that transition.

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm 

And don't forget to sign up for The Deep View daily newsletter. We don’t just cover AI, we decode it. In a world flooded with hype, we deliver sharp, no-nonsense insights to keep you ahead of the curve and help you put AI to work every day: subscribe.thedeepview.com

Creators and Guests

Host

Jason Hiner

Editor-in-Chief of The Deep View

What is The Deep View: Conversations?

From frontier labs and enterprise platforms to emerging startups reshaping entire industries, The Deep View: Conversations podcast interviews the brightest minds and the most influential leaders in AI.

Jason Hiner: So, Tara and Ty, thanks for coming on the Deep View Conversations podcast. I appreciate your time. Why don't we start by having you all introduce yourselves and say what you do at OpenAI.

Tara Seshan: Great. Thanks for having us. Excited to be on here. My name is Tara. I am on the product team at OpenAI. I spend time working on ChatGPT Work and how we think about productivity. Generally, there's lots of new things that we're working on and of course partner closely with Ty.

Ty Geri: Yes. Thanks so much for having us on. I'm Ty. I'm also on the product team at OpenAI. And I work on ChatGPT Work and also proactivity. So, how can ChatGPT be more proactively helpful for you across your work and life?

Jason Hiner: Yes. And it's funny because that started, you know, it's an interesting journey because that started a little bit, at least for me, my first experience with that was in this feature that was only available in Pro, which was the, why am I forgetting the name of it, ChatGPT Pulse. Thank you. ChatGPT Pulse, which when you would come on in the morning would sort of find some new stories for you, would suggest things about the way to organize your day, you know, and things like that. And so you didn't have to just sort of go to the chat button, start thinking about what your question was, like it was sort of prompting you. And then when it came with those things, at the end, it would have some follow-up questions. Would you like me to do X? Would you like to know about Y? You know, things like that. So the journey had started a little bit there, you know, not, I don't know, end of last year. But it's taken a whole leap forward with ChatGPT Work and then scheduled tasks. What about this moment, ChatGPT Work launched in June? You know, what have you been seeing in terms of the ways people are using it? I think that would be a great place to start. Yeah.

Ty Geri: So I'm glad that you you were enjoying Pulse. And I'm also glad that, you know, with scheduled tasks, we were able to bring the same set of features and more to over the order of magnitude more people. Right. So Pulse was available in our pro tier. And now scheduled tasks actually yesterday launched to all users. So literally over a billion users now can start to experience sort of like the magic that you started to experience last year, which actually I think connects really well to what a lot of ChatGPT Work is about, which is taking what a lot of users or like a small subset of users actually, we're experiencing at the frontier in terms of like these new agentic capabilities in Codex. And thinking about, okay, how do we take those capabilities and bring them to literally an order of magnitude, more people and make it more accessible. And that's what ChatGPT Work is trying to do. And what we saw internally was adoption of Codex internally, beyond software engineering, was exponential across finance, across marketing, across legal. And then people are also using it across their personal lives. So it's sort of, you know, these adoption curves are slightly lagging development and research, but not by that much. And the question was like, okay, how do we then take this and bring it outside and bring it to all of our paid tiers from plus and up. And what we've seen is an incredible adoption curve, sort of similar to what we saw internally. I'm sure you've seen like our hockey stick graphs of like people for the first time, a lot of people for the first time, experiencing what it feels like to not just get an answer from ChatGPT, but to actually get something done. And I think that's sort of like the magic we want to compound and really help more people get more things done.

Tara Seshan: Yeah, certainly more people get more things done over increasingly longer periods of time with increased task complexity, like even in the time since we've launched ChatGPT Work, people are not only capable of asking it to do something small and take an action, rather than simply just asking a question, they can actually say, oh, go do this thing. But the size and scale of those things that they're able to ask it to do have gone from something small to, oh, no way to build this feature for me or like, or build this site for me to, oh, wait, now go build this entire experience as tool I might use for like one time thing or a ongoing tool that my team uses to like base all of their operations on. So I think what I personally also get very excited about is like seeing how the product like elevates people's ambitions. As they interact with it, they might start by trying it out to like test it and see something like some small benefit. But over time, we've just seen even an individual user's journeys, like some even people internally, like our finance team, etc. They go from, oh, let me like have it do a very specific thing to, oh, wait, like, you know, huge parts of their job can be both assisted by this product and even shared across many, many team members.

Jason Hiner: Yeah. You touched on the thing that I think is the sort of the biggest question is like, can, you know, OpenAI with the sort of capability overhang of like the models have gotten so much more capable, especially in the past year, can you, can you take that and start to get more people involved in doing things more than just thinking of ChatGPT as a smarter version of Google search, right? Yeah. And that is sort of what ChatGPT Work has set itself out to do. Like, okay, we're going to take all this sort of magic that would happen with like agents and Codex and harnesses. And now like, we're going to try to make this, this is my perception of it. I want you to like make sure I've got it right, that it's going to take that we're going to take it and make it easier for normal people who are just using ChatGPT to start asking it to do some things for it and not just answer some some questions or some queries.

Ty Geri: Yeah, I mean, I think that's exactly spot on. And I think it's not just about sort of easier. It's also just about just for like, quote unquote, normal people, but it's about just more accessible for for everybody, right? Now you have an agent, like a full assistant that has a computer in the cloud that can has a browser, and can do things and connect plugins. And all of this is available, not just on your laptop, but on your phone. Yeah. And and that's, you know, that's sort of like just like this incredible experience that you can suddenly have access to these capabilities from from anywhere. So it's also sort of for more people,

Tara Seshan: But also just from anywhere, which is I think, yeah, it's not just accessible. It's also across more surfaces, and also more powerful. Like as a part of getting ChatGPT Work out there, we added a ton of features to both to work in and Codex to make this thing capable of doing much more stuff for people out of the box easily. And a really like visceral feeling use case of that is that even Codex users whilst using ChatGPT Work on mobile are like, this is the future. This is here. So it really wasn't just about making it accessible to everybody as a part of doing that. You also increase like the power and the quality of that user experience for just about everyone on the product.

Jason Hiner: Okay, so that's the everyone perspective. Now I want to go to the other end of the spectrum like the technical perspective. So it is the same, you know, agent harness ChatGPT Work is the same agent harness that runs Codex. It's the same, you know, technology, the same things. It's just packaged differently. But it also operates technically a little bit differently as well. So talk a little bit about the difference with because when you go into the app, you can sort of switch between I'm talking about the desktop Codex, and then sort of the ChatGPT. And then in ChatGPT, you have sort of like chat and work. And so, but when you're in Codex, you're really off operating more locally. And then when you're in ChatGPT for work, you're operating in the cloud. Is that the best way to like the simplest way to think about it?

Tara Seshan: Yeah. So ultimately, what we want here is for a lot of these nuances and complexities to fall away. So we use it and have to keep this mapping in their mind. Like the thing that we aspire to do and the thing that our product should live up to those aspirations is that a user just comes to the desktop app, for example, and in front of them is the white box, and they can add in their intent in there and we do the right thing for them. Like that is that is truly the intent. ChatGPT Work and Codex are like, you know, using the same agent harness as mentioned. But if you are a developer who uses Codex, and you flipped that, that drop down to Codex, you are not missing out on anything by not flipping over to the other tab. Actually, if you ask in the Codex tab for it to make a PowerPoint deck for you, or to schedule a task or something like that, you are getting the same experience as you would get on ChatGPT Work. Of course, given the task, we are showing the appropriate UI. So when you in the Codex tab ask for it to build a feature for you, you were showing like work trees and the diff pane and all of these things. And when you ask something like build me a PowerPoint deck, we are not showing code relevant UI. But actually, that's simply a view switcher on like what's going on under the hood. As a Codex user, you are basically using ChatGPT Work when you ask like a non technical query. So you don't need to like switch between these products at all. And our aspiration is like that switching no longer is something that users have to cognitively do. You mentioned a little bit about like local versus cloud. Ultimately, the benefit of doing things in the cloud is that exactly that thing we had mentioned before that like I'm on my phone, I can like type in a request I can go on the muni get off and this thing is just done for me and that is an amazing, amazing experience. But there's certainly power in like having the agent have access to like files on my machine as well. And that's for coding that is especially relevant. But again, our aspiration is like, based on your request, how do we figure out the right thing to do for you as a user at the right time and take the like switching and all of those pieces out of it.

Jason Hiner: Yeah. Thank you. Because certainly from a user standpoint, like that makes the most sense. And yeah, so you got at one of what the most challenging things that I've come across and using the products. And this is sort of the difference between, I think someone who maybe used Codex and use ChatGPT, and then like someone who didn't. So I think if you never used Codex, like ChatGPT Work, it's like, I just flip between them or sometimes it can even sort of do it itself, right? I think like you put the right sort of kind of query in and my sense is, if you put it into ChatGPT and ask it to do something, it's sort of like is is kind of like nudging you over to work, you don't have to flip a switch or anything. And so I think the and I know a lot of people that have worked in AI and that have like journalists cover an AI who've never used agents, right? So even still, so I think ChatGPT Work brought that made that like a lot more simple. They're like using agents for the first time. The challenge is if you used Codex before and then you're and so I'll give you this example and this challenge that I had and this is where I don't envy you all with the work you have to do from a UI standpoint because this stuff gets tricky, you know, but this is good. This is the feedback we need. Okay, so here's a here's a scenario where I run into a challenge a couple times. So for our podcast, one of the things we do is we and this was like an unlock when this happened, when I realized like, oh, I could use Codex to just make a transcript of the podcast. I got just and before if I tried to upload that file, like, oh, man, it was is just take forever, right? Because video files are are huge as our friends, you know, here now. And so I realized like, oh, on my local machine, I can have Codex. Here's the file. Yeah, do it. And then I want you to make a transcript. I want you to do like the timestamp that you put in the Google, you know, in sorry, in YouTube description. And it was magical, right? Like, it just did that in like, three minutes, maybe maybe 10 minutes, if it was a really big file. And then when ChatGPT Work came, I like did that. And I was like, Oh, I want to go back. I need to grab that prompt that I did from in Codex. And then I was like, Oh, no, it was in Codex. And I did it locally. And so like, I can't I can't get to that the prompt that I wrote for for Codex, because I did it locally. So so I've had that a couple times, that's one specific example. But I've had a couple times where I did it like Codex on my computer. And then I went to look it up on my phone, I was like, Oh, shoot, I did that on the other sort of side. And so those are the things where I know that gets really tricky between sort of getting between local and cloud.

Ty Geri: Yeah, I mean, I think your your pinpointing exactly a massive pinpoint, which is, you know, you're doing things locally. And then you want to continue them in the cloud on the go. This is something that we are actively working. Okay, like we you know, this is not the experience that users should have. But we want to make sure that you know, as we are sort of continuing on this journey to make things work seamlessly across all our surfaces, that we're still, you know, going back to what you just said about, you know, there's a whole set of people that haven't experienced these authentic experiences, we're not holding anything back, right? So it's sort of like this balance of like, we need to bring all this power to as many people as possible. And at the same time, make sure that these things work all seamlessly together. And there's there's like a, you know, there's a trade off here. And I think we're constantly trying to, you know, close gaps. But at the same time, while you're closing up, you're adding more power. And so it's like, it's this balance. But, but that's, that's an awesome example of exactly the type of syncing between local and cloud that we want to enable for users. So you can do that exactly that transcript experience, and then still get that prompt and continue. And maybe that file is actually stored in your library. And you can continue working on it.

Tara Seshan: I would add that that trade off is not only between like, releasing it as soon as possible to be able to get more people using these products and making it accessible. It is also that until now, a lot of product progress in AI has been local. Like a lot of these initial moments were about like, you do all these things locally. And as such, you unlock a ton of functionality and development like inherently is so local. And so we are trying to actually balance like maybe like three things, which is how do we keep local, which is like worked very effectively continue working very effectively. How do we like reach a whole set of people who haven't tried agents before? And how do we accelerate as much as possible, the like infrastructure work and the development time that we're investing in internally, such that all of these things are cohesive across all of these surfaces. So we want to we're straddling like a three way split here and are really trying to bring it together as effectively as possible.

Jason Hiner: This is where I don't envy your job. Like I said, but but I also I appreciate it because like it's important work and like you have to do it in order to get this to a place where more people are going to be comfortable with it, where a billion people are going to use, you know, these things and eventually more. But continue to hold us accountable. It's important that we make these things seamless. Excellent. Thank you. So I just want to maybe zoom out just a minute to to think about, even if you think about a year ago, right, like last fall, from where we are now to like a year ago, I remember it was about this time last year, where so I worked, I've only been working at the deep view, you know, which is AI focused publications since December. So it was about this time last year, as I'm working at a more general purpose publication where we cover an all of tech, although AI was a specific part of it. And I had the sense then, like, okay, this is starting to move so fast that I'm not keeping up. And I was like, I would really like to do something where I could read about it, think about it, learn about it, talk about it, write about it every day. And then that was before agentic happened, that was before, you know, coding agents, you know, all of these things. So you think back even a year ago, like, we weren't having any conversations about, you know, really agents on, I mean, there was, there was a little bit, but it wasn't, you know, anywhere to the degree of where we are now, like, even I remember when Shook and I would schedule this podcast, when we talked in June, and we're like, oh, yeah, let's schedule it for June. But like, also, who knows what's going to happen? Are we scheduled for August? We're like, who knows what's going to happen by then, right? So my question is, all of this has progressed so fast in 2026, you know, what were, if you all are working specifically on agents now, what were you even working on a year ago? And could you have anticipated, you know, this, what did you anticipate, you know, your work might look like a year from now?

Ty Geri: Yeah, I mean, it is quite incredible to think, you know, what we were working on a year ago, and how far we've sort of come since then. I think like the, I'll start with like the way I work has completely transformed in the last year, from, you know, using Codex, sort of like as a complete daily driver to just get so much leverage out of everything I do, to, you know, the way we work as a team, in terms of like the way we can prioritize things and set deadlines and the ambition we have. And like, I think the way we build products is different. Like, it used to be that we'd spend more time debating, you know, what should this look like, feel like, etc. Yeah. And now we can just build it and feel it and see and my hope is that our products are better because of that, because we're actually able to just experience the actual thing we're building much faster. And these things were just like not possible. It's like almost like the time between imagination and creation has shrunk exponentially over the last 12 months. And I think that makes working on anything more exciting.

Jason Hiner: Is that because you can, you can come up with an idea, say you come up with an idea, instead of before you go and sort of debate it with your team, you can come up with the idea, you can have the agent prototype it for you, then you can bring that prototype and have them look at that versus you sort of going and having a very only sort of theoretical level conversation of something.

Ty Geri: Yeah, which is like, you know, we talk about that as if that's like a thing of the past, but that's like literally how all work was done up until

Tara Seshan: There's a class of questions where they're just better answered by like trying out the thing. And now you can just try out the thing. Like there's certainly discussions that you still need to have regardless, like there's a level of thoughtfulness, like we regardless have to bring to our products and we still do. But I think the there's just some type of question we're like, but this interaction pattern, but this interaction pattern is like, just try it out and see it. Now we just get to try it out and see that part's awesome. I would say the other thing that I feel like is like a way we work that is maybe continue to stay consistent is the way to build is like maybe think two to three months about what the models are going to be capable of and then try to get out of the way of the model as much as possible. And even last year to now, like the big Delta has been the model's capabilities, like as you know, as everyone says constantly, like this is the worst they're ever going to be. And so the biggest change of course in how we're building things is the models are capable of just so much more now. And as such, again, like we're more ambitious about what we can get done, we're more ambitious about the experiences that we want to build, experiences that weren't quite right back then because the models weren't quite right are now possible. And that has been like one of the most exciting things about about working here and in this space is that, yeah, you just wait a little bit and now the models are really great and all these product experiences that you imagine you can now roll out to users, like agents, everyone kept saying, you know, a couple years ago, like this is the year of agents, no, no, no, this is the year of agents. But like last year was the year of agents, and they're only going to get more persistent or capable as time goes on.

Ty Geri: Yeah, I mean, something that I've chatted with a bunch of people about is how, you know, there are so many demos that I've seen over the last three years in the AI space. And if you look back at all those demos, they were all, you know, over a year ago, like things that, you know, were cool, but we're not like enterprise grade or consumer grade. And then if you try any of those demos today, they just work. And like now it's like we just take it for granted, right, that they just work, and they're getting faster and faster. And I think like those improvements in latency are going to be another massive unlock that we'll see over time. But like the reality is like, we're no longer in in demo land for agents. And I think that's like this massive change that has happened almost subtly. But yeah, it's been really amazing.

Jason Hiner: Ty, you mentioned starting your day in Codex. And so it's interesting, more and more people I talked to at OpenAI over the past, even like two months, I said this, like, I don't I don't start with Slack, I don't start with email, I go to Codex. Could you both talk to me a little bit about like, what does that look like for you? And ChatGPT Work, like, where does ChatGPT Work come in to that as well? Like, how do you start how if you're going to start your work, you're going to start your day with the agent? Like, how does that work versus, you know, I still think most people are following their their notifications, right? And that kind of thing. And it sort of pulls your attention in certain directions.

Ty Geri: Yeah, yeah, I mean, so maybe I'll clarify a little bit more. I start my day in ChatGPT Work. But even more than that, I would say ChatGPT Work starts my day before I start my day, because like, I have a scheduled task that runs through everything I've missed in Slack and anything that's blocked on me. That like every day at 8am, the first thing I look at before I open Slack, is I look at that because that actually focuses me because slack and so many things going on. But that's like, this is blocked on me, I missed this message, I need to respond here, this is waiting on this. And honestly, it's such a nice way to start my day, because I already feel after going through that list, that I'm like in control, as opposed to starting in Slack, where I feel like completely out of control. Yeah. And that's literally on my phone. That's like right before I commute to work is that's that's like how I start my day.

Tara Seshan: I do the same thing. And actually, the way I start the one of the most useful things I have is like a self updating to do list. So I have like a little app for myself that I have in Codex that is a to do list of things that I can add, but also like my agent can add things to it. And as I do them, it like ticks it off of like things that I've accomplished and just having like a surface where I can look at that see what I have going on and both of us are kind of like working to get it done is like very, very helpful as a collaborative surface between me and my agent to like see to see like what we have to do for today. I love starting my day in Codex or in ChatGPT Work rather than Slack, because I feel like Slack is so like Slack's always pushing things towards me. And this way I kind of get to pull like what are my highest priorities? What are the things that actually didn't really need my attention? Like what are actually the more like longer enduring things that I have to work on rather than getting constantly distracted by immediate set of asks. And so actually, I feel like a lot of the toil or like immediate response to stuff my agent can just do most of it and just ask me for like approve, approve, approve to get some things done, like send this message or send the summary, etc. Or like update this review doc or something like that. It just does it for me now. And so actually, my time is kind of saved for doing the important things. And I can track all the important things on my to do list. And so that all goes nicely.

Jason Hiner: Okay, Tara, I have to click double click on the list because so I've had this so I've, you know, have it doing a morning, you know, thing for me as well doing a morning. So I've set up this essentially chief of staff agent to go, you know, go through Slack, go through Gmail, go through my calendar. And then I was using Google Tasks. And I was like, and it couldn't access my to do list. So I was like, Oh, what can I do? And then I like asked it, could you write, you know, there are what what could you do to make it so we can access it, right? So I just asked it to do it. And then it created an mcp and all of that. But it still doesn't work like it hasn't actually worked. So what you did it sounds like is you had to make your make it to do list. Is it an app? How did you how did you get it to do your tasks?

Tara Seshan: Honestly, I literally asked it make it to do list app. I want it to be really, really simple. It's literally just marked down in check boxes. Yeah. And it adds things to my list. And it I cross things off my list. It is literally that simple. It is really not that complicated. And I'm like just a huge believer in what I think some people call like the malleable software movement or like the cozy software movement. It goes back to like what Alan Kay said in the 60s about like the true personal computer will only be personal when you have personalized software as well. Like there are things that I want out of my software that are very, very specific. Maybe it's because I'm PM. I don't know. But like I want really specific things out of my software and all of the other like tools to build software, whatever, even like, oh, like the block based wiki editor, like I don't need those. Some of those things are too complicated. I don't want to learn all that stuff. But now with like Codex or with ChatGPT Work, I can like build the little app I really wanted. It doesn't cost me anything to maintain. Like I'm not spending a ton of time and effort. It's exactly suited to the thing that I want to do. It's like being in a kitchen where you had the perfect like single purpose kitchen device for everything you've ever wanted to do. But you don't have the responsibility to manage it all. Like it's an ideal scenario. Or you don't have to store it. You don't have that extra counter space. Exactly. Like imagine you had a kitchen where you could have a single purpose kitchen device for everything you had to do. Like the perfect, you know, lemon zester for your lemon. For the way Tara likes to have lemons. For the way Tara likes to have lemons. But you never had to store it in any drawers. There was no problems with that. You never had to like clean up after. That's kind of how my life is right now with software. It's great.

Jason Hiner: It's a great metaphor. I really like it. So let's talk a little bit about this too. Ty, are there any things that you've had it build for you that are specific to things that you do? I also have an example of one I can share.

Ty Geri: Yeah. I mean, I think there are a couple things. So one, I spend a lot of my time looking at metrics. Okay. And it's actually like, I find it's kind of hard to get to the point of like diminishing returns on understanding how, you know, people are using the product and you know, where are we seeing friction and how can we improve? And so I've worked with it just basically, you know, I've created like a set of skills that run every day that sort of just try to ask like the questions that I would ask every single day about the product and look for anomalies or look for things that we can fix and make better. And that's something that it's like, if we talk about like the sort of like the joy of taking away menial labor versus the joy of like having a real thought partner, I really feel like I'm at the corner of like the joy of a real thought partner. That's at work in my personal life. It's also really fun to make sites. And you can make all sorts of exciting sites like I had a friend that made like an awesome wedding invite site that was like super interactive, like it let you like control either the bride or the groom. And you could like walk around the venue and like visit different stations for information about and like they did it on their phone and made it exactly what they wanted. Like this like malleable, people wanted malleable software, maybe like, you know, malleable wedding invites. It's just it's so much joy.

Jason Hiner: I'll share one that I did. So that I thought was was really interesting. So I have, this is this is when ChatGPT Atlas went away. So one of the actually really cool things about ChatGPT Atlas is it let you save your prompts, you know. And so I thought, okay, ChatGPT Atlas is going away. I had a lot of prompts saved that I use every day or at least multiple times a week, probably like 10 that I use regularly. And so I thought, well, I need to find a way to do it. And so I was using like one of the I was using this product called Raycast, which is essentially a shortcut thing. And I thought, you know what I would really want is I just want like a drop down on the Mac menu that I drop it down. And then it puts them in different categories. And then I can just select that prompt. And then whatever browser, whatever pizza software I do, it can just drop it in there. And I thought, man, I wish there was something like that. And I'm looking around trying to find it. Could I do it in Raycast? It's like, Oh, yeah, there's an API and you can do it. And then I was like, I should just ask, you know, ChatGPT if it could make it for me, right? It made it, it made it in like 15 minutes. And it made exactly what I wanted. It made the thing. And then it was like, and then it had multiple levels of things that it suggested is like, okay, here's phase two. And I was like, okay, what's phase two? And it's like, let's make, you know, a markdown thing, you can save it in someplace like Google Drive, or maybe like Apple, you know, Apple Drive, Apple, so that you could have it shared to from your desktop to your laptop. And I was like, Oh, that's a great idea. Go ahead and do phase two is like 10 minutes. And it did that as well. So, so within like less than an hour, I had this perfect piece of software. It's like this software extension that that like, I'm not going to go have it sell it in the app store or something. It's just exactly for me. It's like the kitchen tool that, you know, that I need. Yeah. And so it's great.

Ty Geri: I actually started doing something this week that I'm amazed that like, I'm continuously every day amazed at what it can do. And I think like it's a lot about like, you know, the more context you bring into tech to be work and Codex, the more capable it is. So, you know, if you connect your email and you connect your calendar, and you can connect even, you know, for, for plus subscribers and up your finances to just be where and it can just like operate across your entire life. And like, I now I asked how to do work every time I book a flight or a hotel or like book anything. I want you to just check every day to see if the price is dropped. And if I can rebook. Yeah. And it does it every day. And it like it checks like, you know, hotel drop, but a little bit, you can rebook, you just rebook. And it's like, it's magic. I mean, it's really having this like personal assistant that is working on your behalf.

Jason Hiner: Very cool. Very cool. All right, I want to get to some one of the challenges of this stuff, because once you start using it, quickly, you can realize like all the things it can do. And then you can also start burning a lot of tokens as well. So I want to ask about this challenge, because I'll give you another example that I had where I had, and I was telling Shoki your the PR on your PR team about this that I was taking something and having it scan something on the web every 10 minutes, pull it, and then put it into a Slack channel for my team so that they didn't have to go and keep like refreshing, you know, a page. It works great. It like it did it. It just made a Slack post every time. And then I switched from like the pro tier to the plus tier, and I like ran out of tokens in like a couple hours, because it's running, you know, so often. And so, you know, agents can be really token heavy, right? And so one of the things, another one of the PR team members I was talking to about this problem on your team, Lauren, shout out to Lauren, which is like, hey, should I do this and then maybe use like one of the lessons? Should I try? Is there a way that I could tell it to use like Luna, because OpenAI recently cut the prices for Luna, made it very reasonable. And so, but then that's a little, I'm still been trying to figure out how to do that, how to tell it, okay, use this model instead, this less expensive model. When people start, you know, working with these things and finding the capabilities, that's the thing that they start running into is like, oh, you can, so that pro tier is $200 a month tier, like the regular plus tier, when I went from pro to plus, all of a sudden I was burning too many tokens, right? I know that's another thing, when you think about usability, you know, you would prefer users not to have to think about that. Like if the, if it could, if it could use the whatever tier you're on and the job might run a little slower, I'm assuming you would love for the model to automatically sort of make that decision. So I'm putting words in your mouth, but like, how would you think about managing that as people start using the product more?

Ty Geri: Yeah, I mean, I think, again, I think you're pointing out like a failure in the product, which is, you know, if you, if you have to have the, if the user has to have the burden of deciding like what model is the best for, you know, which use case and, you know, using this model, that model. And I think this is something that, you know, we're wrestling with. Okay. And it maybe even goes back to that, that tradeoff of like, how do we bring, you know, the most powerful things to as many people as possible, but still, you know, wrestle with this, this issue of, you know, like, how do we make it simple? And so, you know, one thing I'll call out is that our research team is incredible and they continuously push the frontier of performance and efficiency. And so we're continuously able to give more capability at a more affordable cost, really month by month. And if you look back, you know, if we talked about like what's available a year ago, it's like you look at all these features that used to be sort of like higher tier features and suddenly they're more available.

Tara Seshan: The intelligence per dollar per token or sorry, the intelligence per token per dollar has only dramatically gone up. Like the amount of intelligence you can buy per dollar that you spend is, it's crazy. So in some ways, like, you know, our research team has made such a commitment to making that possible and continue to make that possible. Like so that even at the highest tier of intelligence is such an efficient model. When it when it comes to its like token spend,

Ty Geri: And that's something we really, really care about pushing. But I think this user experience of knowing when to use the right model is something that we're thinking a lot about.

Jason Hiner: Okay. And hopefully we'll be making it better. Because like the ideal scenario would be like the interface, the harness, the agent, like makes that decision of like, okay, you're on this tier. And so we're going to use this model, you know, maybe even a model that's sort of this model, the last generation model is powerful enough to do this thing. And the intelligence is a lot less expensive. And so, but it also can't predict what you're going to be doing next week, right? Either. So it's also one of the challenges. But if it knows that like, okay, you're going to burn through all these tokens, you know, in an hour, then it's going to maybe make a choice to move you to a different model. Is that so that's that's part of the that sort of model choice piece. Is that part of the way that you all think about the product to and design the product? That's maybe the bigger question.

Ty Geri: Yeah. I mean, I think we never want to sacrifice like user control. But I think we do want to give a more seamless like happy path for sort of getting things done in the most efficient way possible. And I don't I don't think we've quite cracked that yet. But this is something that we are we're we're thinking about a lot.

Tara Seshan: How do we have the most intelligent defaults possible in the sense that like, based on what you're trying to do, how do we nudge your default you to the right choice while giving users like the control they need if they want to based on their task to override those decisions.

Jason Hiner: Yeah, very good. How about, you know, in March, Greg Brockman, president of OpenAI, talked about the super app, right? The first time I was I remember being in a with a group of journalists and, you know, he even mentioned it. He talked about the super app. And it was like it had sort of been rumored. And then you just sort of admit it. Yes, we're working on a super app, we're going to bring all these things together. It's going to be one thing. But as I understand it, OpenAI has never necessarily declared like, we're at the super app now, right? We don't like

Tara Seshan: I don't like the term super. I don't think that the super term quite makes sense. I think we also talk about that. I think Thibault Sottiaux, of course, has like also said publicly that like, super app is like not really like an internal word that we use to talk about what we're trying to do. It's it's about how do we bring the power of agents to as many people as possible, the over a billion people on on ChatGPT, how do we do so whilst making these like, complex tradeoffs between giving the power users the power to do like the maximum set of things at the frontier of intelligence and capability, how do we bring this and make this as successful as possible, and how do we have the right product experience and infrastructure to like bridge those things like that I think is the maybe very prosaic way of framing this thing that is sounds as exciting than super app. But as to in terms of like a product building guidepost, the frame of like the super app is not as useful versus like, oh, who are we trying to serve? Like, how do we make this possible? How do we, you know, bring the best models and capabilities to our users and like, get out of the way of the model being like really excellent?

Ty Geri: Yeah, I do think though that the like today, the ChatGPT app across mobile web and desktop has access to the full suite of most powerful features built on same agentic harness and powered by our frontier models. And that is, you know, has been sort of a journey over time, making sure that you know, you just come to ChatGPT. And when you're in ChatGPT, we can get you the most powerful and capable experience and personal assistant as possible. I think we're still on that journey, right? And some of the things that you brought up shows that we're challenges, we have a way to go. But I think that the direction, and at least what we can see from adoption of these features, I feel confident that we're in the right direction. Because I think if we, if we weren't, we wouldn't be seeing more and more people using ChatGPT in new ways, which is just, you know, we want to meet our users where they are. And so like the super option be the app that people use every day. I think that's like the way I see it's just like the thing that you open, whether it's desktop, local, like local cloud, mobile web, you shouldn't have to worry that you're you're talking to the right thing or the wrong thing.

Jason Hiner: And it is now one app, maybe that's the most important thing. Because even, you know, four months ago, you're using or I was using ChatGPT Atlas and Codex, you know, now there's only

Tara Seshan: One, there's only one thing. Do you use the browser in the in the desktop app? The Chrome extension.

Jason Hiner: I use the Chrome extension and then just like the in app browser inside the in app browser. Yes, like it's still I've heard people saying that they're using it for a lot of like they're using a lot of their browsing right in the browser. It just there's some things it doesn't do, you know, bookmarks, you know, other things.

Tara Seshan: We have a list. We have a list of many things we're trying to build into it. But but yeah, that is that is the aspiration.

Jason Hiner: I mean, I'm sort of open to that idea of like if you have a big screen and you sort of can have the whole screen there and like maybe on half of it is sort of ChatGPT on one side and the other side of the browser. Like yeah, that would be pretty.

Tara Seshan: Yeah, you can you can do that today and we're making a bunch of improvements to that.

Jason Hiner: All right, I'll keep an eye on that for sure. For sure. How about what's the thing that you can't wait for this to be able to do? You know, there's so many more things that it can do. There's so many more things that it can do from just a few months ago. We talked about a year ago like it's just like something that didn't exist or you know, fully like now is is sort of gaining so many more capabilities all the time. You know, what's the thing that you wish that you can do with ChatGPT Work or you're looking forward to being able to do with ChatGPT Work, you know, broadly. I know we no product announcements.

Tara Seshan: Yeah, no product announcements. I think I think folks have talked about this publicly, but one of the things I'm most excited for is like more persistence. I think that being able to do my work over longer and longer periods where I am increasingly steering it from like a higher and higher level has been amazing. The transition from like auto complete to like have it build a feature to have it build an app to like slash goal. If you've used slash goal to now things running over really long periods to get stuff done has been I love the direction of where that's going. I love that direction for me personally because it makes my work day just so much better and so much more fun. But I also love that transition in terms of like, I love the idea that like work feels like a multiplayer game where I'm working with like my friends and agents to like get stuff done together. And I just feel like that future is like more and more in in our grasp.

Jason Hiner: Tara, can you give me an example of something like a slash goal that you could share that that you've that you've tried?

Tara Seshan: A very recent one. I'm trying to like a recent slash goal that I might have done. Oh, like I'm trying to build a prototype to like express a bunch of ideas of like what I want to do to to some like side nav changes. And so just like slash goal here's like a here's a mock just go do it and like see what happens. Yeah. Okay.

Ty Geri: Yeah, I would I would I would big plus one to persistence. And I think like from a proactivity perspective, the I'm really looking forward to it. And I feel like you even mentioned this before about how we're feeling that the models are getting a little bit more proactive. And the way we're feeling that is you know, maybe through scheduled tasks. Maybe we're feeling that through the model suddenly suggesting things when you open the app on the homepage. But to me, it's just that's just the start. Because like I really think like, you know, if we imagine like our true personal assistant that can help you achieve your goals, both personally and at work, they should be curious and about how they can help and think one step ahead. And like, I'm really excited about the way we interact with these models, becoming more and more proactive, unless just reactive, not just waiting on me to come ask a question. Exactly. And I think, you know, we're starting to do this. So you'll see that in conversations in chat, you can see the model started start to suggest scheduled tasks that can do. Hey, do you want me to monitor this for you? Hey, do you want me to check in this on this in a week from now? And then when you open the app, if you have plugins connected, you can get these personalized, what to do next. But I think we're really at the beginning. And I think, like our amazing research team is working really hard about how do we make the models more proactively helpful for our users. And I think that's going to one, bring down this problem of everybody needing to figure out, okay, what can I use these models to do to help me out? And it could be more about the model saying like, Hey, here's how I can help you.

Tara Seshan: Can I give you another one? I'm really excited about it's this is like less, you know, sort of like, aspirationally high minded, it's really prosaic, which is I just want it to work with every single third party tool I use. Like your Google task thing kills me. I like wanted to work with Google tasks. I want all those things to work. I want them all to be connected and those connections to be excellent and really reliable and fantastic. Like, these things are as good as like all the data that you give it such that it can perform all these actions in your life. Computer use is sick. Browser use is sick. All those things are great. I still want it to work with Google tasks out of the box. And so like, it can't work with every single system, whether those are like really old school ones, like whatever, I don't know, my like UCSF wants me to log into to manage my patient care all the way to like really cool new things, like awesome apps that we use like like linear, for example, it's just work with everything. Just work with all your tools. It's just work. And of course, some of that is is things we can do to our plugin platform. We really care about developers on that platform. Other things are like third party tools just like builds into our platform. It's really easy now with Codex and they can do it really easily, especially in Codex, like we've made a developing plugins really good in Codex itself. Oh, nice. And so yeah, I just wanted to work with all my third party tools. It sounds like you wanted to work with them.

Jason Hiner: I do for sure. And like multiple accounts too, like I want to work Gmail and my personal Gmail.

Tara Seshan: Yes. What are your other feature requests? What do you want? I want to know.

Jason Hiner: Yeah, very good. Yeah, those are those are great. My other one would be and you sort of touch on this a little bit. The when you go to ChatGPT now and you go to ChatGPT Work, like there'll be like three things where it says, you know, that proactive things suggest things like I would love for that to keep getting like more and more personal of like, Hey, and like reference maybe my my daily my morning sort of thing and say, you still haven't done this one yet. Do you are you ready to send that email or something like that? I would love for to get like even more specific to me. I know that's, you know, challenging, but that could be really I think that would be so powerful. It's like you still haven't done this thing yet. I, like, go to ChatGPT Work to do it and I'm like, Oh, yeah, shoot. I need to yes. All I need to do is approve that right click it a hit approve and then it sends the email for me to finish something. You know, that would be amazing.

Tara Seshan: I love it. You mentioned earlier sorry to now ask you questions, but you had mentioned earlier that there are lots of journalists or folks who even maybe cover AI or technology who haven't used agents yet. What would get them to use agents?

Jason Hiner: I know I know this this is true. It's so interesting, right? Like not just journalists, but like people that that are like working enterprise that have used AI for years and like all you know, the question I'm always asking people like so what agents are you using? What are you having it do? Things like that. And and they're like, Yeah, I'm not really into age. I haven't really tried agents yet. This does get to actually one of the questions I had when I asked them like what are the what are what's holding you back? You know, what do you want to see? Most of the obstacle actually comes down to like privacy and security. Yeah, they have reservations about connecting their Gmail or their Slack or that kind of thing. So I I'd love to hear like what it what do you all say to people that have sort of those reservations of like well is now my private data going to be used to train models all of those kinds of things.

Tara Seshan: There is like a whole set of things that we do for enterprise that are like classic enterprise guarantees about how we use their data. We've always had like a policy around this like every enterprise who buys ChatGPT like we have that's like agreement with them. Certainly for folks on on our business plan as well. Like there's there's a set of like classic enterprise things like where's your data stored like what are all the like security restrictions are like the ACLs around who you can view and share information that you have here. What plugins could be connected? There's also just a deep investment we've made in admins and like what control and visibility and observability they have over their organization, the data, how it flows, actions that are taken like that whole suite. Some of that is just bread and butter things that should just be like excellent about any enterprise product. And so like I think that the more interesting things that we can do are how do we enable enterprises like for some of these less like like deterministic things like this person has access or this person doesn't have access but how can we help them evaluate some of these like questions or these like like data sharing or even security questions using the models themselves. We have some really amazing like cyber models we have really amazing folks who think about this internally. The models are increasingly good at being able to ask and understand these questions and so I'm not only excited about there are lots of like core enterprise things that we have done and continue to invest in but like how can we also build products here that help solve this problem for enterprises in new ways.

Ty Geri: Yeah and I think also outside of the enterprise lens what's so important is that users are always in control and you know I'd say for a user that wants to connect for example their email they have full control over you know different ways ChatGPT can interact with that plugin. ChatGPT can be read only and it can have also ask always right before it does anything with any of your external plugins it will ask you and I think like these are these options and control exist precisely to make it so that the users are at the steering wheel and that they're able to experience these capabilities but at the at the sort of level of comfort that they have and hopefully over time you know build trust with the system as you see that oh you know it's asking the right questions and you know we're not diving too deep and so I think like a big part about this you know both in the enterprise with admins and both for all of our self-serve tiers is like how can we make sure that the users always feel in control and that's like really how we're building the product

Jason Hiner: And is it is it fair to say like on that personal data and stuff for the people who have concerns about you know oh it's going to you know train on my personal data and then somebody's going to be able to sort of reverse engineer the prompt and figure out you know my my private data you know that's something that you feel comfortable you know sharing like that's you don't have to worry

Ty Geri: Yeah there are multiple levels of sort of security and privacy and an organization and again in terms of control users are completely in control over whether or not any data they share with ChatGPT can be used to train other models it's fully in and user control in the settings

Jason Hiner: Yes. Yeah, yeah. Very good, Tara and Ty. We can keep doing this, I know, for a lot longer. I really appreciate your time. It's been great learning more about ChatGPT Work. I know the audience is going to be really excited to learn it, and I hope that from this more of them will give it a try, because there's a lot of stuff that it can help them do.

Tara Seshan: If we can tell them to do one thing—like, go pick a task, file or resolve your parking ticket with it, or go make a site. You made a site, it sounds like, and really love sites. Like, go make a site. Are there other things you just tell them to try out of the box?

Ty Geri: Yeah, I mean, I think if you can try computer use on the desktop experience, it just blows your mind that it can check into a flight for you. Simple things like triage email—and, you know, do pretty much anything that you would do on your computer. I think, to me, if you want to experience the frontier, computer use is where it's at.

Jason Hiner: And if you're just getting started, have it go through your email for you and pick what are the things that I haven't responded to that need a response from me. Things like that are great ones to start with, or—

Ty Geri: Delete all the spam in my email.

Tara Seshan: Yeah, yeah. Tell me any of these spam. I like Ty's hotel rebooking one. That's a really good one. I'm going to do that.

Jason Hiner: Yeah, you should definitely do that. That sounds awesome. Very good. Thank you both. Good luck as you keep making the product better.

Tara Seshan: Thank you. Send us your feature requests.

Ty Geri: Keep sharing the feedback. It's good. Thank you.